macOS Β· Local-First Β· Private by Architecture

AI that runs
on your machine.
Not theirs.

CortexDesk is a local desktop assistant for macOS teams that need speed and privacy. It understands natural language, uses local tools, and asks for approval before sensitive actions.

No cloud. No subscriptions. No compromise.

0ms
network latency
0B
data transmitted
6+
specialized micro-models
∞
ownership of your context
// 00 β€” what is cortexdesk

What CortexDesk is β€” and what it is not.

CortexDesk is designed for professionals, not for autonomous internet agents. It focuses on local execution, explicit approvals, and predictable outcomes.

What it is
A local assistant for real desktop work.

You can ask in natural language and CortexDesk routes to local tools for calendar, reminders, drafting, search, and reasoning support.

  • Runs on your Mac
  • Typed + voice workflows
  • Action previews before execution
What it is not
Not a remote-control bot for everything.

CortexDesk does not aim to automate every website and app without guardrails. It is not an always-on cloud operator.

  • No default cloud dependency
  • No hidden background actions
  • No blind autonomy on sensitive tasks
Current status
Early product, practical scope.

V1 is focused: local assistant loop, approvals, memory, calendar/reminders, voice, and specialist routing. Office server and full remote automation come later.

  • Built first for macOS
  • Stable core before expansion
  • Roadmap-driven releases
// 01 β€” pipeline

How the orchestration works

Every request passes through a coordinated pipeline of specialized models β€” all running locally, all under your control.

01 β€”
User Input

You type a command, ask a question, or delegate a task. CortexDesk accepts natural language or structured directives.

Local only
02 β€”
Cortex Router

The orchestration core classifies intent and routes your request to the right micro-model combination β€” no guessing, no bloat.

On-device inference
03 β€”
Micro-Model Execution

Specialized models handle their domain in parallel: safety screening, writing assistance, calendar logic, numerical reasoning.

Parallel inference
04 β€”
Approval Gate

Before any action executes β€” sending email, editing a file, accessing calendar β€” you see exactly what will happen and approve it.

Explicit approval
// 02 β€” micro-models

Specialized models.
Coordinated precisely.

Rather than one monolithic model doing everything, CortexDesk routes tasks to purpose-built micro-models β€” each optimized, lightweight, and isolated.

CORTEX ORCHESTRATOR
The coordination core

The central routing layer that classifies, decomposes, and delegates. Runs entirely in-process. No syscalls leave the orchestration sandbox without an approval token.

cortex-core v0.9 Β· on-device
β†’ routes to specialized agents
safety
Guard

Pre-screens all inputs. Flags policy conflicts before execution. Cannot be bypassed by prompts or roleplay.

scheduling
Chrono

Reads and writes calendar events locally via CalendarKit. Understands time, conflicts, and context.

writing
Scribe

Drafts, edits, and formats documents. Adapts to your voice, your contacts, your context.

code
Forge

Writes and refactors code. Runs sandboxed in local execution with no outbound access.

math
Calculus

Symbolic and numeric reasoning for financial and legal work. Deterministic. No LLM guessing.

memory
Recall

Manages your local context store. Retrieves selectively. Encrypted at rest. Always yours.

// 03 β€” why local

The case for local-first.

For professionals who handle client data, financial records, or privileged communications β€” the cloud is not an option.

Privacy
Zero data egress.
Total confidentiality.

Client data, case notes, financial models β€” none of it ever touches a remote server. Privilege is preserved. Compliance is structural, not promised.

  • Attorney-client privilege protected
  • HIPAA / SOC 2 compatible by design
  • No telemetry, no analytics, no logging
  • Air-gap deployable
Speed
No round-trips.
Instant response.

Local inference means sub-100ms first token on Apple Silicon. No queuing, no rate limits, no degraded service during outages. Your machine is the server.

  • Runs on M1/M2/M3/M4 natively
  • Neural Engine acceleration
  • Offline-capable, always available
  • No throttling or tier limits
Ownership
One purchase.
Perpetual license.

No monthly subscription. No per-seat pricing. No API costs that scale with usage. Pay once, run forever. Your workflow is never held hostage to a pricing model.

  • Perpetual license, one-time fee
  • Updates included for 2 years
  • No vendor lock-in on your data
  • Source escrow for enterprise
// 04 β€” control
CortexDesk β€” Action Queue
β–Ά cortex run "draft memo for Q4 board review"
β†’ Routing to: Scribe + Calculus
β†’ Reading: /Documents/Financials/Q4-2024.xlsx
β†’ Draft complete. 847 words. Proposing action:
PENDING APPROVAL β€” write to /Documents/Memos/
Save memo "Q4-board-review-draft.docx" to /Documents/Memos/ ?

You approve every action.

CortexDesk never acts autonomously. Before any file is modified, email sent, or system call made β€” you see a plain-language summary and choose to allow or deny.

Every proposed action is legible.
No hidden background processes.
No silent writes.
CortexDesk tells you exactly what it wants to do β€” and waits.

// ready to run local

Built for professionals
who can't afford to trust the cloud.

Join the waitlist. Early access ships to legal, finance, and developer teams first.

macOS 13+ Β· Apple Silicon native Β· Ships Q3 2026